src/HOL/Tools/Sledgehammer/sledgehammer_filter.ML
author blanchet
Thu May 12 15:29:19 2011 +0200 (2011-05-12 ago)
changeset 42728 44cd74a419ce
parent 42702 d7c127478ee1
child 42729 e011f632227c
permissions -rw-r--r--
added configuration options for experimental features
blanchet@38988
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_filter.ML
blanchet@38027
     2
    Author:     Jia Meng, Cambridge University Computer Laboratory and NICTA
blanchet@36393
     3
    Author:     Jasmin Blanchette, TU Muenchen
blanchet@39958
     4
blanchet@39958
     5
Sledgehammer's relevance filter.
wenzelm@33309
     6
*)
paulson@15452
     7
blanchet@38988
     8
signature SLEDGEHAMMER_FILTER =
wenzelm@16802
     9
sig
blanchet@38993
    10
  datatype locality = General | Intro | Elim | Simp | Local | Assum | Chained
blanchet@38752
    11
blanchet@40070
    12
  type relevance_fudge =
blanchet@41159
    13
    {allow_ext : bool,
blanchet@41790
    14
     local_const_multiplier : real,
blanchet@41159
    15
     worse_irrel_freq : real,
blanchet@40070
    16
     higher_order_irrel_weight : real,
blanchet@40070
    17
     abs_rel_weight : real,
blanchet@40070
    18
     abs_irrel_weight : real,
blanchet@40070
    19
     skolem_irrel_weight : real,
blanchet@40070
    20
     theory_const_rel_weight : real,
blanchet@40070
    21
     theory_const_irrel_weight : real,
blanchet@40070
    22
     intro_bonus : real,
blanchet@40070
    23
     elim_bonus : real,
blanchet@40070
    24
     simp_bonus : real,
blanchet@40070
    25
     local_bonus : real,
blanchet@40070
    26
     assum_bonus : real,
blanchet@40070
    27
     chained_bonus : real,
blanchet@40070
    28
     max_imperfect : real,
blanchet@40070
    29
     max_imperfect_exp : real,
blanchet@40070
    30
     threshold_divisor : real,
blanchet@40070
    31
     ridiculous_threshold : real}
blanchet@40070
    32
blanchet@35966
    33
  type relevance_override =
blanchet@40070
    34
    {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
    35
     del : (Facts.ref * Attrib.src list) list,
blanchet@40070
    36
     only : bool}
blanchet@35966
    37
blanchet@42646
    38
  val trace : bool Config.T
blanchet@42728
    39
  val new_monomorphizer : bool Config.T
blanchet@42728
    40
  val ignore_no_atp : bool Config.T
blanchet@42728
    41
  val instantiate_inducts : bool Config.T
blanchet@42680
    42
  val is_locality_global : locality -> bool
blanchet@40205
    43
  val fact_from_ref :
blanchet@38996
    44
    Proof.context -> unit Symtab.table -> thm list
blanchet@38996
    45
    -> Facts.ref * Attrib.src list -> ((string * locality) * thm) list
blanchet@41767
    46
  val all_facts :
blanchet@42641
    47
    Proof.context -> 'a Symtab.table -> bool -> thm list -> thm list
blanchet@42641
    48
    -> (((unit -> string) * locality) * (bool * thm)) list
blanchet@41768
    49
  val const_names_in_fact :
blanchet@41768
    50
    theory -> (string * typ -> term list -> bool * term list) -> term
blanchet@41768
    51
    -> string list
blanchet@42638
    52
  val is_dangerous_term : term -> bool
blanchet@37347
    53
  val relevant_facts :
blanchet@42638
    54
    Proof.context -> real * real -> int
blanchet@41336
    55
    -> (string * typ -> term list -> bool * term list) -> relevance_fudge
blanchet@41066
    56
    -> relevance_override -> thm list -> term list -> term
blanchet@40070
    57
    -> ((string * locality) * thm) list
paulson@15347
    58
end;
paulson@15347
    59
blanchet@38988
    60
structure Sledgehammer_Filter : SLEDGEHAMMER_FILTER =
paulson@15347
    61
struct
paulson@15347
    62
blanchet@38652
    63
open Sledgehammer_Util
blanchet@38652
    64
blanchet@42646
    65
val trace =
blanchet@42646
    66
  Attrib.setup_config_bool @{binding sledgehammer_filter_trace} (K false)
blanchet@42646
    67
fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
blanchet@35826
    68
blanchet@41273
    69
(* experimental features *)
blanchet@42728
    70
val new_monomorphizer =
blanchet@42728
    71
  Attrib.setup_config_bool @{binding sledgehammer_new_monomorphizer} (K false)
blanchet@42728
    72
val ignore_no_atp =
blanchet@42728
    73
  Attrib.setup_config_bool @{binding sledgehammer_ignore_no_atp} (K false)
blanchet@42728
    74
val instantiate_inducts =
blanchet@42728
    75
  Attrib.setup_config_bool @{binding sledgehammer_instantiate_inducts} (K false)
blanchet@38827
    76
blanchet@38993
    77
datatype locality = General | Intro | Elim | Simp | Local | Assum | Chained
blanchet@38752
    78
blanchet@42449
    79
(* (quasi-)underapproximation of the truth *)
blanchet@42680
    80
fun is_locality_global Local = false
blanchet@42680
    81
  | is_locality_global Assum = false
blanchet@42680
    82
  | is_locality_global Chained = false
blanchet@42680
    83
  | is_locality_global _ = true
blanchet@42449
    84
blanchet@40070
    85
type relevance_fudge =
blanchet@41159
    86
  {allow_ext : bool,
blanchet@41790
    87
   local_const_multiplier : real,
blanchet@41159
    88
   worse_irrel_freq : real,
blanchet@40070
    89
   higher_order_irrel_weight : real,
blanchet@40070
    90
   abs_rel_weight : real,
blanchet@40070
    91
   abs_irrel_weight : real,
blanchet@40070
    92
   skolem_irrel_weight : real,
blanchet@40070
    93
   theory_const_rel_weight : real,
blanchet@40070
    94
   theory_const_irrel_weight : real,
blanchet@40070
    95
   intro_bonus : real,
blanchet@40070
    96
   elim_bonus : real,
blanchet@40070
    97
   simp_bonus : real,
blanchet@40070
    98
   local_bonus : real,
blanchet@40070
    99
   assum_bonus : real,
blanchet@40070
   100
   chained_bonus : real,
blanchet@40070
   101
   max_imperfect : real,
blanchet@40070
   102
   max_imperfect_exp : real,
blanchet@40070
   103
   threshold_divisor : real,
blanchet@40070
   104
   ridiculous_threshold : real}
blanchet@40070
   105
blanchet@35966
   106
type relevance_override =
blanchet@40070
   107
  {add : (Facts.ref * Attrib.src list) list,
blanchet@40070
   108
   del : (Facts.ref * Attrib.src list) list,
blanchet@40070
   109
   only : bool}
paulson@21070
   110
blanchet@37616
   111
val sledgehammer_prefix = "Sledgehammer" ^ Long_Name.separator
blanchet@39896
   112
val abs_name = sledgehammer_prefix ^ "abs"
blanchet@39896
   113
val skolem_prefix = sledgehammer_prefix ^ "sko"
blanchet@38992
   114
val theory_const_suffix = Long_Name.separator ^ " 1"
blanchet@37616
   115
blanchet@40227
   116
fun needs_quoting reserved s =
blanchet@40375
   117
  Symtab.defined reserved s orelse
wenzelm@42290
   118
  exists (not o Lexicon.is_identifier) (Long_Name.explode s)
blanchet@40227
   119
blanchet@40375
   120
fun make_name reserved multi j name =
blanchet@40227
   121
  (name |> needs_quoting reserved name ? quote) ^
wenzelm@41491
   122
  (if multi then "(" ^ string_of_int j ^ ")" else "")
blanchet@38744
   123
blanchet@40375
   124
fun explode_interval _ (Facts.FromTo (i, j)) = i upto j
blanchet@40375
   125
  | explode_interval max (Facts.From i) = i upto i + max - 1
blanchet@40375
   126
  | explode_interval _ (Facts.Single i) = [i]
blanchet@40375
   127
blanchet@41279
   128
val backquote =
blanchet@41279
   129
  raw_explode #> map (fn "`" => "\\`" | s => s) #> implode #> enclose "`" "`"
blanchet@40205
   130
fun fact_from_ref ctxt reserved chained_ths (xthm as (xref, args)) =
blanchet@38744
   131
  let
blanchet@38996
   132
    val ths = Attrib.eval_thms ctxt [xthm]
blanchet@38996
   133
    val bracket =
blanchet@41999
   134
      map (enclose "[" "]" o Pretty.str_of o Args.pretty_src ctxt) args
blanchet@41999
   135
      |> implode
blanchet@40375
   136
    fun nth_name j =
blanchet@38996
   137
      case xref of
blanchet@41279
   138
        Facts.Fact s => backquote s ^ bracket
blanchet@38996
   139
      | Facts.Named (("", _), _) => "[" ^ bracket ^ "]"
blanchet@40375
   140
      | Facts.Named ((name, _), NONE) =>
blanchet@40375
   141
        make_name reserved (length ths > 1) (j + 1) name ^ bracket
blanchet@40375
   142
      | Facts.Named ((name, _), SOME intervals) =>
blanchet@40375
   143
        make_name reserved true
blanchet@40375
   144
                 (nth (maps (explode_interval (length ths)) intervals) j) name ^
blanchet@40375
   145
        bracket
blanchet@38744
   146
  in
blanchet@40375
   147
    (ths, (0, []))
blanchet@38752
   148
    |-> fold (fn th => fn (j, rest) =>
blanchet@40375
   149
                 (j + 1, ((nth_name j,
blanchet@38752
   150
                          if member Thm.eq_thm chained_ths th then Chained
blanchet@38752
   151
                          else General), th) :: rest))
blanchet@38744
   152
    |> snd
blanchet@38699
   153
  end
blanchet@37616
   154
blanchet@41199
   155
(* This is a terrible hack. Free variables are sometimes code as "M__" when they
blanchet@41199
   156
   are displayed as "M" and we want to avoid clashes with these. But sometimes
blanchet@41199
   157
   it's even worse: "Ma__" encodes "M". So we simply reserve all prefixes of all
blanchet@41199
   158
   free variables. In the worse case scenario, where the fact won't be resolved
blanchet@41199
   159
   correctly, the user can fix it manually, e.g., by naming the fact in
blanchet@41199
   160
   question. Ideally we would need nothing of it, but backticks just don't work
blanchet@41199
   161
   with schematic variables. *)
blanchet@41199
   162
fun all_prefixes_of s =
blanchet@41199
   163
  map (fn i => String.extract (s, 0, SOME i)) (1 upto size s - 1)
blanchet@41199
   164
fun close_form t =
blanchet@41199
   165
  (t, [] |> Term.add_free_names t |> maps all_prefixes_of)
blanchet@41199
   166
  |> fold (fn ((s, i), T) => fn (t', taken) =>
blanchet@41199
   167
              let val s' = Name.variant taken s in
blanchet@41199
   168
                ((if fastype_of t' = HOLogic.boolT then HOLogic.all_const
blanchet@41199
   169
                  else Term.all) T
blanchet@41199
   170
                 $ Abs (s', T, abstract_over (Var ((s, i), T), t')),
blanchet@41199
   171
                 s' :: taken)
blanchet@41199
   172
              end)
blanchet@41199
   173
          (Term.add_vars t [] |> sort_wrt (fst o fst))
blanchet@41199
   174
  |> fst
blanchet@41199
   175
blanchet@41199
   176
fun string_for_term ctxt t =
blanchet@41199
   177
  Print_Mode.setmp (filter (curry (op =) Symbol.xsymbolsN)
blanchet@41199
   178
                   (print_mode_value ())) (Syntax.string_of_term ctxt) t
blanchet@41199
   179
  |> String.translate (fn c => if Char.isPrint c then str c else "")
blanchet@41199
   180
  |> simplify_spaces
blanchet@41199
   181
blanchet@41199
   182
(** Structural induction rules **)
blanchet@41199
   183
blanchet@41200
   184
fun struct_induct_rule_on th =
blanchet@41199
   185
  case Logic.strip_horn (prop_of th) of
blanchet@41199
   186
    (prems, @{const Trueprop}
blanchet@41199
   187
            $ ((p as Var ((p_name, 0), _)) $ (a as Var (_, ind_T)))) =>
blanchet@41199
   188
    if not (is_TVar ind_T) andalso length prems > 1 andalso
blanchet@41199
   189
       exists (exists_subterm (curry (op aconv) p)) prems andalso
blanchet@41199
   190
       not (exists (exists_subterm (curry (op aconv) a)) prems) then
blanchet@41199
   191
      SOME (p_name, ind_T)
blanchet@41199
   192
    else
blanchet@41199
   193
      NONE
blanchet@41199
   194
  | _ => NONE
blanchet@41199
   195
blanchet@41207
   196
fun instantiate_induct_rule ctxt concl_prop p_name ((name, loc), (multi, th))
blanchet@41199
   197
                            ind_x =
blanchet@41199
   198
  let
blanchet@41199
   199
    fun varify_noninducts (t as Free (s, T)) =
blanchet@41199
   200
        if (s, T) = ind_x orelse can dest_funT T then t else Var ((s, 0), T)
blanchet@41199
   201
      | varify_noninducts t = t
blanchet@41199
   202
    val p_inst =
blanchet@41199
   203
      concl_prop |> map_aterms varify_noninducts |> close_form
blanchet@41199
   204
                 |> lambda (Free ind_x)
blanchet@41199
   205
                 |> string_for_term ctxt
blanchet@41199
   206
  in
blanchet@41207
   207
    ((fn () => name () ^ "[where " ^ p_name ^ " = " ^ quote p_inst ^ "]", loc),
blanchet@41199
   208
     (multi, th |> read_instantiate ctxt [((p_name, 0), p_inst)]))
blanchet@41199
   209
  end
blanchet@41199
   210
blanchet@41199
   211
fun type_match thy (T1, T2) =
blanchet@41199
   212
  (Sign.typ_match thy (T2, T1) Vartab.empty; true)
blanchet@41199
   213
  handle Type.TYPE_MATCH => false
blanchet@41199
   214
blanchet@41199
   215
fun instantiate_if_induct_rule ctxt stmt stmt_xs (ax as (_, (_, th))) =
blanchet@41200
   216
  case struct_induct_rule_on th of
blanchet@41199
   217
    SOME (p_name, ind_T) =>
wenzelm@42361
   218
    let val thy = Proof_Context.theory_of ctxt in
blanchet@41199
   219
      stmt_xs |> filter (fn (_, T) => type_match thy (T, ind_T))
blanchet@41199
   220
              |> map_filter (try (instantiate_induct_rule ctxt stmt p_name ax))
blanchet@41199
   221
    end
blanchet@41199
   222
  | NONE => [ax]
blanchet@41199
   223
wenzelm@28477
   224
(***************************************************************)
wenzelm@28477
   225
(* Relevance Filtering                                         *)
wenzelm@28477
   226
(***************************************************************)
mengj@19194
   227
paulson@24287
   228
(*** constants with types ***)
paulson@24287
   229
blanchet@38939
   230
fun order_of_type (Type (@{type_name fun}, [T1, @{typ bool}])) =
blanchet@38939
   231
    order_of_type T1 (* cheat: pretend sets are first-order *)
blanchet@38939
   232
  | order_of_type (Type (@{type_name fun}, [T1, T2])) =
blanchet@38939
   233
    Int.max (order_of_type T1 + 1, order_of_type T2)
blanchet@38939
   234
  | order_of_type (Type (_, Ts)) = fold (Integer.max o order_of_type) Ts 0
blanchet@38939
   235
  | order_of_type _ = 0
blanchet@38939
   236
blanchet@38823
   237
(* An abstraction of Isabelle types and first-order terms *)
blanchet@38823
   238
datatype pattern = PVar | PApp of string * pattern list
blanchet@38939
   239
datatype ptype = PType of int * pattern list
blanchet@38744
   240
blanchet@38823
   241
fun string_for_pattern PVar = "_"
blanchet@38823
   242
  | string_for_pattern (PApp (s, ps)) =
blanchet@38823
   243
    if null ps then s else s ^ string_for_patterns ps
blanchet@38823
   244
and string_for_patterns ps = "(" ^ commas (map string_for_pattern ps) ^ ")"
blanchet@38939
   245
fun string_for_ptype (PType (_, ps)) = string_for_patterns ps
paulson@24287
   246
paulson@24287
   247
(*Is the second type an instance of the first one?*)
blanchet@38824
   248
fun match_pattern (PVar, _) = true
blanchet@38824
   249
  | match_pattern (PApp _, PVar) = false
blanchet@38824
   250
  | match_pattern (PApp (s, ps), PApp (t, qs)) =
blanchet@38824
   251
    s = t andalso match_patterns (ps, qs)
blanchet@38824
   252
and match_patterns (_, []) = true
blanchet@38824
   253
  | match_patterns ([], _) = false
blanchet@38824
   254
  | match_patterns (p :: ps, q :: qs) =
blanchet@38824
   255
    match_pattern (p, q) andalso match_patterns (ps, qs)
blanchet@38939
   256
fun match_ptype (PType (_, ps), PType (_, qs)) = match_patterns (ps, qs)
paulson@24287
   257
blanchet@38823
   258
(* Is there a unifiable constant? *)
blanchet@38827
   259
fun pconst_mem f consts (s, ps) =
blanchet@38939
   260
  exists (curry (match_ptype o f) ps)
blanchet@38827
   261
         (map snd (filter (curry (op =) s o fst) consts))
blanchet@38827
   262
fun pconst_hyper_mem f const_tab (s, ps) =
blanchet@38939
   263
  exists (curry (match_ptype o f) ps) (these (Symtab.lookup const_tab s))
blanchet@37505
   264
blanchet@38939
   265
fun pattern_for_type (Type (s, Ts)) = PApp (s, map pattern_for_type Ts)
blanchet@38939
   266
  | pattern_for_type (TFree (s, _)) = PApp (s, [])
blanchet@38939
   267
  | pattern_for_type (TVar _) = PVar
blanchet@38827
   268
blanchet@38744
   269
(* Pairs a constant with the list of its type instantiations. *)
blanchet@41204
   270
fun ptype thy const x =
blanchet@38939
   271
  (if const then map pattern_for_type (these (try (Sign.const_typargs thy) x))
blanchet@41204
   272
   else [])
blanchet@41204
   273
fun rich_ptype thy const (s, T) =
blanchet@41204
   274
  PType (order_of_type T, ptype thy const (s, T))
blanchet@41204
   275
fun rich_pconst thy const (s, T) = (s, rich_ptype thy const (s, T))
paulson@24287
   276
blanchet@38939
   277
fun string_for_hyper_pconst (s, ps) =
blanchet@38939
   278
  s ^ "{" ^ commas (map string_for_ptype ps) ^ "}"
paulson@24287
   279
blanchet@38823
   280
(* Add a pconstant to the table, but a [] entry means a standard
blanchet@38819
   281
   connective, which we ignore.*)
blanchet@41066
   282
fun add_pconst_to_table also_skolem (s, p) =
blanchet@41066
   283
  if (not also_skolem andalso String.isPrefix skolem_prefix s) then I
blanchet@41066
   284
  else Symtab.map_default (s, [p]) (insert (op =) p)
blanchet@38819
   285
blanchet@40369
   286
fun pconsts_in_terms thy is_built_in_const also_skolems pos ts =
blanchet@37505
   287
  let
blanchet@38819
   288
    val flip = Option.map not
blanchet@38587
   289
    (* We include free variables, as well as constants, to handle locales. For
blanchet@41205
   290
       each quantifiers that must necessarily be skolemized by the automatic
blanchet@41205
   291
       prover, we introduce a fresh constant to simulate the effect of
blanchet@41205
   292
       Skolemization. *)
blanchet@40373
   293
    fun do_const const x ts =
blanchet@41336
   294
      let val (built_in, ts) = is_built_in_const x ts in
blanchet@41336
   295
        (not built_in
blanchet@41336
   296
         ? add_pconst_to_table also_skolems (rich_pconst thy const x))
blanchet@41336
   297
        #> fold do_term ts
blanchet@41336
   298
      end
blanchet@38827
   299
    and do_term t =
blanchet@38827
   300
      case strip_comb t of
blanchet@38827
   301
        (Const x, ts) => do_const true x ts
blanchet@38827
   302
      | (Free x, ts) => do_const false x ts
blanchet@38939
   303
      | (Abs (_, T, t'), ts) =>
blanchet@38939
   304
        (null ts
blanchet@41066
   305
         ? add_pconst_to_table true (abs_name, PType (order_of_type T + 1, [])))
blanchet@38827
   306
        #> fold do_term (t' :: ts)
blanchet@38827
   307
      | (_, ts) => fold do_term ts
blanchet@38939
   308
    fun do_quantifier will_surely_be_skolemized abs_T body_t =
blanchet@37537
   309
      do_formula pos body_t
blanchet@38747
   310
      #> (if also_skolems andalso will_surely_be_skolemized then
blanchet@41066
   311
            add_pconst_to_table true
blanchet@41207
   312
                (gensym skolem_prefix, PType (order_of_type abs_T, []))
blanchet@38587
   313
          else
blanchet@38587
   314
            I)
blanchet@38587
   315
    and do_term_or_formula T =
blanchet@41273
   316
      if T = HOLogic.boolT then do_formula NONE else do_term
blanchet@37537
   317
    and do_formula pos t =
blanchet@37537
   318
      case t of
blanchet@38939
   319
        Const (@{const_name all}, _) $ Abs (_, T, t') =>
blanchet@38939
   320
        do_quantifier (pos = SOME false) T t'
blanchet@37537
   321
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   322
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   323
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   324
        fold (do_term_or_formula T) [t1, t2]
blanchet@37537
   325
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@41140
   326
      | @{const False} => I
blanchet@41140
   327
      | @{const True} => I
blanchet@37537
   328
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@38939
   329
      | Const (@{const_name All}, _) $ Abs (_, T, t') =>
blanchet@38939
   330
        do_quantifier (pos = SOME false) T t'
blanchet@38939
   331
      | Const (@{const_name Ex}, _) $ Abs (_, T, t') =>
blanchet@38939
   332
        do_quantifier (pos = SOME true) T t'
haftmann@38795
   333
      | @{const HOL.conj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38795
   334
      | @{const HOL.disj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38786
   335
      | @{const HOL.implies} $ t1 $ t2 =>
blanchet@37537
   336
        do_formula (flip pos) t1 #> do_formula pos t2
haftmann@38864
   337
      | Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@38587
   338
        fold (do_term_or_formula T) [t1, t2]
blanchet@38587
   339
      | Const (@{const_name If}, Type (_, [_, Type (_, [T, _])]))
blanchet@38587
   340
        $ t1 $ t2 $ t3 =>
blanchet@38587
   341
        do_formula NONE t1 #> fold (do_term_or_formula T) [t2, t3]
blanchet@38939
   342
      | Const (@{const_name Ex1}, _) $ Abs (_, T, t') =>
blanchet@38939
   343
        do_quantifier (is_some pos) T t'
blanchet@38939
   344
      | Const (@{const_name Ball}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   345
        do_quantifier (pos = SOME false) T
blanchet@38939
   346
                      (HOLogic.mk_imp (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@38939
   347
      | Const (@{const_name Bex}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   348
        do_quantifier (pos = SOME true) T
blanchet@38939
   349
                      (HOLogic.mk_conj (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@37537
   350
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@37537
   351
        do_term t0 #> do_formula pos t1  (* theory constant *)
blanchet@37537
   352
      | _ => do_term t
blanchet@38819
   353
  in Symtab.empty |> fold (do_formula pos) ts end
paulson@24287
   354
paulson@24287
   355
(*Inserts a dummy "constant" referring to the theory name, so that relevance
paulson@24287
   356
  takes the given theory into account.*)
blanchet@41200
   357
fun theory_constify ({theory_const_rel_weight, theory_const_irrel_weight, ...}
blanchet@41200
   358
                     : relevance_fudge) thy_name t =
blanchet@40070
   359
  if exists (curry (op <) 0.0) [theory_const_rel_weight,
blanchet@40070
   360
                                theory_const_irrel_weight] then
blanchet@41200
   361
    Const (thy_name ^ theory_const_suffix, @{typ bool}) $ t
blanchet@37505
   362
  else
blanchet@41200
   363
    t
blanchet@41200
   364
blanchet@41200
   365
fun theory_const_prop_of fudge th =
blanchet@41200
   366
  theory_constify fudge (Context.theory_name (theory_of_thm th)) (prop_of th)
blanchet@37505
   367
paulson@24287
   368
(**** Constant / Type Frequencies ****)
paulson@24287
   369
blanchet@38743
   370
(* A two-dimensional symbol table counts frequencies of constants. It's keyed
blanchet@38743
   371
   first by constant name and second by its list of type instantiations. For the
blanchet@38823
   372
   latter, we need a linear ordering on "pattern list". *)
paulson@24287
   373
blanchet@38823
   374
fun pattern_ord p =
blanchet@38743
   375
  case p of
blanchet@38744
   376
    (PVar, PVar) => EQUAL
blanchet@38823
   377
  | (PVar, PApp _) => LESS
blanchet@38823
   378
  | (PApp _, PVar) => GREATER
blanchet@38823
   379
  | (PApp q1, PApp q2) =>
blanchet@38823
   380
    prod_ord fast_string_ord (dict_ord pattern_ord) (q1, q2)
blanchet@38939
   381
fun ptype_ord (PType p, PType q) =
blanchet@38939
   382
  prod_ord (dict_ord pattern_ord) int_ord (swap p, swap q)
paulson@24287
   383
blanchet@38939
   384
structure PType_Tab = Table(type key = ptype val ord = ptype_ord)
paulson@24287
   385
blanchet@40204
   386
fun count_fact_consts thy fudge =
blanchet@37503
   387
  let
blanchet@38827
   388
    fun do_const const (s, T) ts =
blanchet@38827
   389
      (* Two-dimensional table update. Constant maps to types maps to count. *)
blanchet@41204
   390
      PType_Tab.map_default (rich_ptype thy const (s, T), 0) (Integer.add 1)
blanchet@38939
   391
      |> Symtab.map_default (s, PType_Tab.empty)
blanchet@38827
   392
      #> fold do_term ts
blanchet@38827
   393
    and do_term t =
blanchet@38827
   394
      case strip_comb t of
blanchet@38827
   395
        (Const x, ts) => do_const true x ts
blanchet@38827
   396
      | (Free x, ts) => do_const false x ts
blanchet@38827
   397
      | (Abs (_, _, t'), ts) => fold do_term (t' :: ts)
blanchet@38827
   398
      | (_, ts) => fold do_term ts
blanchet@40070
   399
  in do_term o theory_const_prop_of fudge o snd end
paulson@24287
   400
paulson@24287
   401
paulson@24287
   402
(**** Actual Filtering Code ****)
paulson@24287
   403
blanchet@39367
   404
fun pow_int _ 0 = 1.0
blanchet@38939
   405
  | pow_int x 1 = x
blanchet@38939
   406
  | pow_int x n = if n > 0 then x * pow_int x (n - 1) else pow_int x (n + 1) / x
blanchet@38939
   407
paulson@24287
   408
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@38824
   409
fun pconst_freq match const_tab (c, ps) =
blanchet@38939
   410
  PType_Tab.fold (fn (qs, m) => match (ps, qs) ? Integer.add m)
blanchet@38939
   411
                 (the (Symtab.lookup const_tab c)) 0
blanchet@38686
   412
paulson@24287
   413
blanchet@38085
   414
(* A surprising number of theorems contain only a few significant constants.
blanchet@38085
   415
   These include all induction rules, and other general theorems. *)
blanchet@37503
   416
blanchet@37503
   417
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@38938
   418
   frequencies. Rare constants give more points if they are relevant than less
blanchet@38938
   419
   rare ones. *)
blanchet@39367
   420
fun rel_weight_for _ freq = 1.0 + 2.0 / Math.ln (Real.fromInt freq + 1.0)
blanchet@38938
   421
blanchet@38938
   422
(* Irrelevant constants are treated differently. We associate lower penalties to
blanchet@38938
   423
   very rare constants and very common ones -- the former because they can't
blanchet@38938
   424
   lead to the inclusion of too many new facts, and the latter because they are
blanchet@38938
   425
   so common as to be of little interest. *)
blanchet@40070
   426
fun irrel_weight_for ({worse_irrel_freq, higher_order_irrel_weight, ...}
blanchet@40070
   427
                      : relevance_fudge) order freq =
blanchet@40070
   428
  let val (k, x) = worse_irrel_freq |> `Real.ceil in
blanchet@38939
   429
    (if freq < k then Math.ln (Real.fromInt (freq + 1)) / Math.ln x
blanchet@38939
   430
     else rel_weight_for order freq / rel_weight_for order k)
blanchet@40070
   431
    * pow_int higher_order_irrel_weight (order - 1)
blanchet@38938
   432
  end
blanchet@37503
   433
blanchet@41790
   434
fun multiplier_for_const_name local_const_multiplier s =
blanchet@41790
   435
  if String.isSubstring "." s then 1.0 else local_const_multiplier
blanchet@38821
   436
blanchet@41790
   437
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@41790
   438
fun generic_pconst_weight local_const_multiplier abs_weight skolem_weight
blanchet@41790
   439
                          theory_const_weight weight_for f const_tab
blanchet@41790
   440
                          (c as (s, PType (m, _))) =
blanchet@41790
   441
  if s = abs_name then
blanchet@41790
   442
    abs_weight
blanchet@41790
   443
  else if String.isPrefix skolem_prefix s then
blanchet@41790
   444
    skolem_weight
blanchet@41790
   445
  else if String.isSuffix theory_const_suffix s then
blanchet@41790
   446
    theory_const_weight
blanchet@41790
   447
  else
blanchet@41790
   448
    multiplier_for_const_name local_const_multiplier s
blanchet@41790
   449
    * weight_for m (pconst_freq (match_ptype o f) const_tab c)
blanchet@41790
   450
blanchet@41790
   451
fun rel_pconst_weight ({local_const_multiplier, abs_rel_weight,
blanchet@41790
   452
                        theory_const_rel_weight, ...} : relevance_fudge)
blanchet@41790
   453
                      const_tab =
blanchet@41790
   454
  generic_pconst_weight local_const_multiplier abs_rel_weight 0.0
blanchet@41790
   455
                        theory_const_rel_weight rel_weight_for I const_tab
blanchet@41790
   456
fun irrel_pconst_weight (fudge as {local_const_multiplier, abs_irrel_weight,
blanchet@41790
   457
                                   skolem_irrel_weight,
blanchet@40070
   458
                                   theory_const_irrel_weight, ...}) const_tab =
blanchet@41790
   459
  generic_pconst_weight local_const_multiplier abs_irrel_weight
blanchet@41790
   460
                        skolem_irrel_weight theory_const_irrel_weight
blanchet@41790
   461
                        (irrel_weight_for fudge) swap const_tab
paulson@24287
   462
blanchet@40070
   463
fun locality_bonus (_ : relevance_fudge) General = 0.0
blanchet@40070
   464
  | locality_bonus {intro_bonus, ...} Intro = intro_bonus
blanchet@40070
   465
  | locality_bonus {elim_bonus, ...} Elim = elim_bonus
blanchet@40070
   466
  | locality_bonus {simp_bonus, ...} Simp = simp_bonus
blanchet@40070
   467
  | locality_bonus {local_bonus, ...} Local = local_bonus
blanchet@40070
   468
  | locality_bonus {assum_bonus, ...} Assum = assum_bonus
blanchet@40070
   469
  | locality_bonus {chained_bonus, ...} Chained = chained_bonus
blanchet@38751
   470
blanchet@40418
   471
fun is_odd_const_name s =
blanchet@40418
   472
  s = abs_name orelse String.isPrefix skolem_prefix s orelse
blanchet@40418
   473
  String.isSuffix theory_const_suffix s
blanchet@40418
   474
blanchet@40204
   475
fun fact_weight fudge loc const_tab relevant_consts fact_consts =
blanchet@40204
   476
  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
blanchet@40204
   477
                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
blanchet@38827
   478
    ([], _) => 0.0
blanchet@38744
   479
  | (rel, irrel) =>
blanchet@40418
   480
    if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
blanchet@40371
   481
      0.0
blanchet@40371
   482
    else
blanchet@40371
   483
      let
blanchet@40371
   484
        val irrel = irrel |> filter_out (pconst_mem swap rel)
blanchet@40371
   485
        val rel_weight =
blanchet@40371
   486
          0.0 |> fold (curry (op +) o rel_pconst_weight fudge const_tab) rel
blanchet@40371
   487
        val irrel_weight =
blanchet@40371
   488
          ~ (locality_bonus fudge loc)
blanchet@40371
   489
          |> fold (curry (op +) o irrel_pconst_weight fudge const_tab) irrel
blanchet@40371
   490
        val res = rel_weight / (rel_weight + irrel_weight)
blanchet@40371
   491
      in if Real.isFinite res then res else 0.0 end
blanchet@38747
   492
blanchet@40369
   493
fun pconsts_in_fact thy is_built_in_const t =
blanchet@38825
   494
  Symtab.fold (fn (s, pss) => fold (cons o pair s) pss)
blanchet@40369
   495
              (pconsts_in_terms thy is_built_in_const true (SOME true) [t]) []
blanchet@40369
   496
fun pair_consts_fact thy is_built_in_const fudge fact =
blanchet@40204
   497
  case fact |> snd |> theory_const_prop_of fudge
blanchet@40369
   498
            |> pconsts_in_fact thy is_built_in_const of
blanchet@38827
   499
    [] => NONE
blanchet@40204
   500
  | consts => SOME ((fact, consts), NONE)
paulson@24287
   501
blanchet@41768
   502
val const_names_in_fact = map fst ooo pconsts_in_fact
blanchet@41768
   503
blanchet@38699
   504
type annotated_thm =
blanchet@38939
   505
  (((unit -> string) * locality) * thm) * (string * ptype) list
blanchet@37505
   506
blanchet@42646
   507
fun take_most_relevant ctxt max_relevant remaining_max
blanchet@42728
   508
        ({max_imperfect, max_imperfect_exp, ...} : relevance_fudge)
blanchet@40070
   509
        (candidates : (annotated_thm * real) list) =
blanchet@38744
   510
  let
blanchet@38747
   511
    val max_imperfect =
blanchet@40070
   512
      Real.ceil (Math.pow (max_imperfect,
blanchet@38904
   513
                    Math.pow (Real.fromInt remaining_max
blanchet@40070
   514
                              / Real.fromInt max_relevant, max_imperfect_exp)))
blanchet@38747
   515
    val (perfect, imperfect) =
blanchet@38889
   516
      candidates |> sort (Real.compare o swap o pairself snd)
blanchet@38889
   517
                 |> take_prefix (fn (_, w) => w > 0.99999)
blanchet@38747
   518
    val ((accepts, more_rejects), rejects) =
blanchet@38747
   519
      chop max_imperfect imperfect |>> append perfect |>> chop remaining_max
blanchet@38744
   520
  in
blanchet@42646
   521
    trace_msg ctxt (fn () =>
wenzelm@41491
   522
        "Actually passed (" ^ string_of_int (length accepts) ^ " of " ^
wenzelm@41491
   523
        string_of_int (length candidates) ^ "): " ^
blanchet@38889
   524
        (accepts |> map (fn ((((name, _), _), _), weight) =>
blanchet@38752
   525
                            name () ^ " [" ^ Real.toString weight ^ "]")
blanchet@38745
   526
                 |> commas));
blanchet@38747
   527
    (accepts, more_rejects @ rejects)
blanchet@38744
   528
  end
paulson@24287
   529
blanchet@40369
   530
fun if_empty_replace_with_locality thy is_built_in_const facts loc tab =
blanchet@38819
   531
  if Symtab.is_empty tab then
blanchet@40369
   532
    pconsts_in_terms thy is_built_in_const false (SOME false)
blanchet@38819
   533
        (map_filter (fn ((_, loc'), th) =>
blanchet@40204
   534
                        if loc' = loc then SOME (prop_of th) else NONE) facts)
blanchet@38819
   535
  else
blanchet@38819
   536
    tab
blanchet@38819
   537
blanchet@42702
   538
fun consider_arities is_built_in_const th =
blanchet@41158
   539
  let
blanchet@41158
   540
    fun aux _ _ NONE = NONE
blanchet@41158
   541
      | aux t args (SOME tab) =
blanchet@41158
   542
        case t of
blanchet@41158
   543
          t1 $ t2 => SOME tab |> aux t1 (t2 :: args) |> aux t2 []
blanchet@41158
   544
        | Const (x as (s, _)) =>
blanchet@41336
   545
          (if is_built_in_const x args |> fst then
blanchet@41158
   546
             SOME tab
blanchet@41158
   547
           else case Symtab.lookup tab s of
blanchet@41158
   548
             NONE => SOME (Symtab.update (s, length args) tab)
blanchet@41158
   549
           | SOME n => if n = length args then SOME tab else NONE)
blanchet@41158
   550
        | _ => SOME tab
blanchet@41158
   551
  in aux (prop_of th) [] end
blanchet@41158
   552
blanchet@42702
   553
(* FIXME: This is currently only useful for polymorphic type systems. *)
blanchet@42702
   554
fun could_benefit_from_ext is_built_in_const facts =
blanchet@42702
   555
  fold (consider_arities is_built_in_const o snd) facts (SOME Symtab.empty)
blanchet@41158
   556
  |> is_none
blanchet@41158
   557
blanchet@40369
   558
fun relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@40070
   559
        (fudge as {threshold_divisor, ridiculous_threshold, ...})
blanchet@40204
   560
        ({add, del, ...} : relevance_override) facts goal_ts =
blanchet@38739
   561
  let
wenzelm@42361
   562
    val thy = Proof_Context.theory_of ctxt
blanchet@40204
   563
    val const_tab = fold (count_fact_consts thy fudge) facts Symtab.empty
blanchet@38819
   564
    val goal_const_tab =
blanchet@40369
   565
      pconsts_in_terms thy is_built_in_const false (SOME false) goal_ts
blanchet@40369
   566
      |> fold (if_empty_replace_with_locality thy is_built_in_const facts)
blanchet@38993
   567
              [Chained, Assum, Local]
blanchet@39012
   568
    val add_ths = Attrib.eval_thms ctxt add
blanchet@39012
   569
    val del_ths = Attrib.eval_thms ctxt del
blanchet@40204
   570
    val facts = facts |> filter_out (member Thm.eq_thm del_ths o snd)
blanchet@38747
   571
    fun iter j remaining_max threshold rel_const_tab hopeless hopeful =
blanchet@38739
   572
      let
blanchet@40191
   573
        fun relevant [] _ [] =
blanchet@38747
   574
            (* Nothing has been added this iteration. *)
blanchet@40070
   575
            if j = 0 andalso threshold >= ridiculous_threshold then
blanchet@38747
   576
              (* First iteration? Try again. *)
blanchet@40070
   577
              iter 0 max_relevant (threshold / threshold_divisor) rel_const_tab
blanchet@38747
   578
                   hopeless hopeful
blanchet@38744
   579
            else
blanchet@40191
   580
              []
blanchet@38889
   581
          | relevant candidates rejects [] =
blanchet@38739
   582
            let
blanchet@38747
   583
              val (accepts, more_rejects) =
blanchet@42646
   584
                take_most_relevant ctxt max_relevant remaining_max fudge
blanchet@42646
   585
                                   candidates
blanchet@38739
   586
              val rel_const_tab' =
blanchet@38745
   587
                rel_const_tab
blanchet@41066
   588
                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
blanchet@38744
   589
              fun is_dirty (c, _) =
blanchet@38744
   590
                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
blanchet@38745
   591
              val (hopeful_rejects, hopeless_rejects) =
blanchet@38745
   592
                 (rejects @ hopeless, ([], []))
blanchet@38745
   593
                 |-> fold (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   594
                              if exists is_dirty consts then
blanchet@38745
   595
                                apfst (cons (ax, NONE))
blanchet@38745
   596
                              else
blanchet@38745
   597
                                apsnd (cons (ax, old_weight)))
blanchet@38745
   598
                 |>> append (more_rejects
blanchet@38745
   599
                             |> map (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   600
                                        (ax, if exists is_dirty consts then NONE
blanchet@38745
   601
                                             else SOME old_weight)))
blanchet@38747
   602
              val threshold =
blanchet@38822
   603
                1.0 - (1.0 - threshold)
blanchet@38822
   604
                      * Math.pow (decay, Real.fromInt (length accepts))
blanchet@38747
   605
              val remaining_max = remaining_max - length accepts
blanchet@38739
   606
            in
blanchet@42646
   607
              trace_msg ctxt (fn () => "New or updated constants: " ^
blanchet@38744
   608
                  commas (rel_const_tab' |> Symtab.dest
blanchet@38822
   609
                          |> subtract (op =) (rel_const_tab |> Symtab.dest)
blanchet@38827
   610
                          |> map string_for_hyper_pconst));
blanchet@38745
   611
              map (fst o fst) accepts @
blanchet@38747
   612
              (if remaining_max = 0 then
blanchet@40191
   613
                 []
blanchet@38745
   614
               else
blanchet@38747
   615
                 iter (j + 1) remaining_max threshold rel_const_tab'
blanchet@38747
   616
                      hopeless_rejects hopeful_rejects)
blanchet@38739
   617
            end
blanchet@38889
   618
          | relevant candidates rejects
blanchet@40204
   619
                     (((ax as (((_, loc), _), fact_consts)), cached_weight)
blanchet@38747
   620
                      :: hopeful) =
blanchet@38739
   621
            let
blanchet@38739
   622
              val weight =
blanchet@38739
   623
                case cached_weight of
blanchet@38739
   624
                  SOME w => w
blanchet@40204
   625
                | NONE => fact_weight fudge loc const_tab rel_const_tab
blanchet@40204
   626
                                      fact_consts
blanchet@38739
   627
            in
blanchet@38741
   628
              if weight >= threshold then
blanchet@38889
   629
                relevant ((ax, weight) :: candidates) rejects hopeful
blanchet@38739
   630
              else
blanchet@38889
   631
                relevant candidates ((ax, weight) :: rejects) hopeful
blanchet@38739
   632
            end
blanchet@38739
   633
        in
blanchet@42646
   634
          trace_msg ctxt (fn () =>
blanchet@38744
   635
              "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
blanchet@38744
   636
              Real.toString threshold ^ ", constants: " ^
blanchet@38744
   637
              commas (rel_const_tab |> Symtab.dest
blanchet@38744
   638
                      |> filter (curry (op <>) [] o snd)
blanchet@38827
   639
                      |> map string_for_hyper_pconst));
blanchet@38889
   640
          relevant [] [] hopeful
blanchet@38739
   641
        end
blanchet@41158
   642
    fun add_facts ths accepts =
blanchet@41167
   643
      (facts |> filter (member Thm.eq_thm ths o snd)) @
blanchet@41167
   644
      (accepts |> filter_out (member Thm.eq_thm ths o snd))
blanchet@40408
   645
      |> take max_relevant
blanchet@38739
   646
  in
blanchet@40369
   647
    facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
blanchet@40204
   648
          |> iter 0 max_relevant threshold0 goal_const_tab []
blanchet@41158
   649
          |> not (null add_ths) ? add_facts add_ths
blanchet@41158
   650
          |> (fn accepts =>
blanchet@42702
   651
                 accepts |> could_benefit_from_ext is_built_in_const accepts
blanchet@41158
   652
                            ? add_facts @{thms ext})
blanchet@42646
   653
          |> tap (fn accepts => trace_msg ctxt (fn () =>
wenzelm@41491
   654
                      "Total relevant: " ^ string_of_int (length accepts)))
blanchet@38739
   655
  end
paulson@24287
   656
blanchet@38744
   657
paulson@24287
   658
(***************************************************************)
mengj@19768
   659
(* Retrieving and filtering lemmas                             *)
mengj@19768
   660
(***************************************************************)
mengj@19768
   661
paulson@33022
   662
(*** retrieve lemmas and filter them ***)
mengj@19768
   663
paulson@20757
   664
(*Reject theorems with names like "List.filter.filter_list_def" or
paulson@21690
   665
  "Accessible_Part.acc.defs", as these are definitions arising from packages.*)
paulson@20757
   666
fun is_package_def a =
blanchet@40205
   667
  let val names = Long_Name.explode a in
blanchet@40205
   668
    (length names > 2 andalso not (hd names = "local") andalso
blanchet@40205
   669
     String.isSuffix "_def" a) orelse String.isSuffix "_defs" a
blanchet@40205
   670
  end
paulson@20757
   671
blanchet@42641
   672
fun mk_fact_table g f xs =
blanchet@42641
   673
  fold (Termtab.update o `(g o prop_of o f)) xs Termtab.empty
blanchet@42641
   674
fun uniquify xs = Termtab.fold (cons o snd) (mk_fact_table I snd xs) []
mengj@19768
   675
blanchet@37626
   676
(* FIXME: put other record thms here, or declare as "no_atp" *)
blanchet@42728
   677
fun multi_base_blacklist ctxt =
blanchet@41199
   678
  ["defs", "select_defs", "update_defs", "split", "splits", "split_asm",
blanchet@41199
   679
   "cases", "ext_cases", "eq.simps", "eq.refl", "nchotomy", "case_cong",
blanchet@41199
   680
   "weak_case_cong"]
blanchet@42728
   681
  |> not (Config.get ctxt instantiate_inducts) ? append ["induct", "inducts"]
blanchet@38682
   682
  |> map (prefix ".")
blanchet@37626
   683
blanchet@37626
   684
val max_lambda_nesting = 3
blanchet@37626
   685
blanchet@37626
   686
fun term_has_too_many_lambdas max (t1 $ t2) =
blanchet@37626
   687
    exists (term_has_too_many_lambdas max) [t1, t2]
blanchet@37626
   688
  | term_has_too_many_lambdas max (Abs (_, _, t)) =
blanchet@37626
   689
    max = 0 orelse term_has_too_many_lambdas (max - 1) t
blanchet@37626
   690
  | term_has_too_many_lambdas _ _ = false
blanchet@37626
   691
blanchet@37626
   692
(* Don't count nested lambdas at the level of formulas, since they are
blanchet@37626
   693
   quantifiers. *)
blanchet@37626
   694
fun formula_has_too_many_lambdas Ts (Abs (_, T, t)) =
blanchet@37626
   695
    formula_has_too_many_lambdas (T :: Ts) t
blanchet@37626
   696
  | formula_has_too_many_lambdas Ts t =
blanchet@41273
   697
    if member (op =) [HOLogic.boolT, propT] (fastype_of1 (Ts, t)) then
blanchet@37626
   698
      exists (formula_has_too_many_lambdas Ts) (#2 (strip_comb t))
blanchet@37626
   699
    else
blanchet@37626
   700
      term_has_too_many_lambdas max_lambda_nesting t
blanchet@37626
   701
blanchet@38692
   702
(* The max apply depth of any "metis" call in "Metis_Examples" (on 2007-10-31)
blanchet@37626
   703
   was 11. *)
blanchet@37626
   704
val max_apply_depth = 15
blanchet@37626
   705
blanchet@37626
   706
fun apply_depth (f $ t) = Int.max (apply_depth f, apply_depth t + 1)
blanchet@37626
   707
  | apply_depth (Abs (_, _, t)) = apply_depth t
blanchet@37626
   708
  | apply_depth _ = 0
blanchet@37626
   709
blanchet@37626
   710
fun is_formula_too_complex t =
blanchet@38085
   711
  apply_depth t > max_apply_depth orelse formula_has_too_many_lambdas [] t
blanchet@37626
   712
blanchet@39946
   713
(* FIXME: Extend to "Meson" and "Metis" *)
blanchet@37543
   714
val exists_sledgehammer_const =
blanchet@37626
   715
  exists_Const (fn (s, _) => String.isPrefix sledgehammer_prefix s)
blanchet@37626
   716
blanchet@38904
   717
(* FIXME: make more reliable *)
blanchet@38904
   718
val exists_low_level_class_const =
blanchet@38904
   719
  exists_Const (fn (s, _) =>
blanchet@38904
   720
     String.isSubstring (Long_Name.separator ^ "class" ^ Long_Name.separator) s)
blanchet@38904
   721
blanchet@38821
   722
fun is_metastrange_theorem th =
blanchet@37626
   723
  case head_of (concl_of th) of
blanchet@37626
   724
      Const (a, _) => (a <> @{const_name Trueprop} andalso
blanchet@37626
   725
                       a <> @{const_name "=="})
blanchet@37626
   726
    | _ => false
blanchet@37626
   727
blanchet@38821
   728
fun is_that_fact th =
blanchet@38821
   729
  String.isSuffix (Long_Name.separator ^ Obtain.thatN) (Thm.get_name_hint th)
blanchet@38821
   730
  andalso exists_subterm (fn Free (s, _) => s = Name.skolem Auto_Bind.thesisN
blanchet@38821
   731
                           | _ => false) (prop_of th)
blanchet@38821
   732
blanchet@37626
   733
val type_has_top_sort =
blanchet@37626
   734
  exists_subtype (fn TFree (_, []) => true | TVar (_, []) => true | _ => false)
blanchet@37626
   735
blanchet@38085
   736
(**** Predicates to detect unwanted facts (prolific or likely to cause
blanchet@37347
   737
      unsoundness) ****)
paulson@21470
   738
blanchet@38289
   739
(* Too general means, positive equality literal with a variable X as one
blanchet@38289
   740
   operand, when X does not occur properly in the other operand. This rules out
blanchet@38289
   741
   clearly inconsistent facts such as X = a | X = b, though it by no means
blanchet@38289
   742
   guarantees soundness. *)
paulson@21470
   743
blanchet@38289
   744
(* Unwanted equalities are those between a (bound or schematic) variable that
blanchet@38289
   745
   does not properly occur in the second operand. *)
blanchet@38607
   746
val is_exhaustive_finite =
blanchet@38607
   747
  let
blanchet@38629
   748
    fun is_bad_equal (Var z) t =
blanchet@38629
   749
        not (exists_subterm (fn Var z' => z = z' | _ => false) t)
blanchet@38629
   750
      | is_bad_equal (Bound j) t = not (loose_bvar1 (t, j))
blanchet@38629
   751
      | is_bad_equal _ _ = false
blanchet@38629
   752
    fun do_equals t1 t2 = is_bad_equal t1 t2 orelse is_bad_equal t2 t1
blanchet@38607
   753
    fun do_formula pos t =
blanchet@38607
   754
      case (pos, t) of
blanchet@38615
   755
        (_, @{const Trueprop} $ t1) => do_formula pos t1
blanchet@38607
   756
      | (true, Const (@{const_name all}, _) $ Abs (_, _, t')) =>
blanchet@38607
   757
        do_formula pos t'
blanchet@38607
   758
      | (true, Const (@{const_name All}, _) $ Abs (_, _, t')) =>
blanchet@38607
   759
        do_formula pos t'
blanchet@38607
   760
      | (false, Const (@{const_name Ex}, _) $ Abs (_, _, t')) =>
blanchet@38607
   761
        do_formula pos t'
blanchet@38607
   762
      | (_, @{const "==>"} $ t1 $ t2) =>
blanchet@38629
   763
        do_formula (not pos) t1 andalso
blanchet@38629
   764
        (t2 = @{prop False} orelse do_formula pos t2)
haftmann@38786
   765
      | (_, @{const HOL.implies} $ t1 $ t2) =>
blanchet@38629
   766
        do_formula (not pos) t1 andalso
blanchet@38629
   767
        (t2 = @{const False} orelse do_formula pos t2)
blanchet@38607
   768
      | (_, @{const Not} $ t1) => do_formula (not pos) t1
haftmann@38795
   769
      | (true, @{const HOL.disj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38795
   770
      | (false, @{const HOL.conj} $ t1 $ t2) => forall (do_formula pos) [t1, t2]
haftmann@38864
   771
      | (true, Const (@{const_name HOL.eq}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   772
      | (true, Const (@{const_name "=="}, _) $ t1 $ t2) => do_equals t1 t2
blanchet@38607
   773
      | _ => false
blanchet@38607
   774
  in do_formula true end
blanchet@38607
   775
blanchet@38592
   776
fun has_bound_or_var_of_type tycons =
blanchet@38592
   777
  exists_subterm (fn Var (_, Type (s, _)) => member (op =) tycons s
blanchet@38592
   778
                   | Abs (_, Type (s, _), _) => member (op =) tycons s
blanchet@38592
   779
                   | _ => false)
paulson@21431
   780
blanchet@38085
   781
(* Facts are forbidden to contain variables of these types. The typical reason
blanchet@37347
   782
   is that they lead to unsoundness. Note that "unit" satisfies numerous
blanchet@38085
   783
   equations like "?x = ()". The resulting clauses will have no type constraint,
blanchet@37347
   784
   yielding false proofs. Even "bool" leads to many unsound proofs, though only
blanchet@37347
   785
   for higher-order problems. *)
blanchet@42638
   786
val dangerous_types = [@{type_name unit}, @{type_name bool}, @{type_name prop}]
paulson@22217
   787
blanchet@38085
   788
(* Facts containing variables of type "unit" or "bool" or of the form
blanchet@38290
   789
   "ALL x. x = A | x = B | x = C" are likely to lead to unsound proofs if types
blanchet@38290
   790
   are omitted. *)
blanchet@42638
   791
val is_dangerous_term =
blanchet@42638
   792
  transform_elim_term
blanchet@42671
   793
  #> (has_bound_or_var_of_type dangerous_types orf is_exhaustive_finite)
paulson@21470
   794
blanchet@42638
   795
fun is_theorem_bad_for_atps thm =
blanchet@38627
   796
  let val t = prop_of thm in
blanchet@38627
   797
    is_formula_too_complex t orelse exists_type type_has_top_sort t orelse
blanchet@42638
   798
    exists_sledgehammer_const t orelse exists_low_level_class_const t orelse
blanchet@42638
   799
    is_metastrange_theorem thm orelse is_that_fact thm
blanchet@38627
   800
  end
blanchet@38627
   801
blanchet@42641
   802
fun meta_equify (@{const Trueprop}
blanchet@42641
   803
                 $ (Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2)) =
blanchet@42641
   804
    Const (@{const_name "=="}, T --> T --> @{typ prop}) $ t1 $ t2
blanchet@42641
   805
  | meta_equify t = t
blanchet@42641
   806
blanchet@42641
   807
val normalize_simp_prop =
blanchet@42641
   808
  meta_equify
blanchet@42641
   809
  #> map_aterms (fn Var ((s, _), T) => Var ((s, 0), T) | t => t)
blanchet@42641
   810
  #> map_types (map_type_tvar (fn ((s, _), S) => TVar ((s, 0), S)))
blanchet@42641
   811
blanchet@38937
   812
fun clasimpset_rules_of ctxt =
blanchet@38937
   813
  let
blanchet@38937
   814
    val {safeIs, safeEs, hazIs, hazEs, ...} = ctxt |> claset_of |> rep_cs
blanchet@38937
   815
    val intros = safeIs @ hazIs
blanchet@38937
   816
    val elims = map Classical.classical_rule (safeEs @ hazEs)
blanchet@42641
   817
    val simps = ctxt |> simpset_of |> dest_ss |> #simps
blanchet@42641
   818
  in
blanchet@42641
   819
    (mk_fact_table I I intros,
blanchet@42641
   820
     mk_fact_table I I elims,
blanchet@42641
   821
     mk_fact_table normalize_simp_prop snd simps)
blanchet@42641
   822
  end
blanchet@38937
   823
blanchet@42641
   824
fun all_facts ctxt reserved really_all add_ths chained_ths =
blanchet@38627
   825
  let
wenzelm@42361
   826
    val thy = Proof_Context.theory_of ctxt
wenzelm@39557
   827
    val global_facts = Global_Theory.facts_of thy
wenzelm@42361
   828
    val local_facts = Proof_Context.facts_of ctxt
blanchet@38644
   829
    val named_locals = local_facts |> Facts.dest_static []
blanchet@38993
   830
    val assms = Assumption.all_assms_of ctxt
blanchet@38993
   831
    fun is_assum th = exists (fn ct => prop_of th aconv term_of ct) assms
blanchet@38752
   832
    val is_chained = member Thm.eq_thm chained_ths
blanchet@42641
   833
    val (intros, elims, simps) = clasimpset_rules_of ctxt
blanchet@42641
   834
    fun locality_of_theorem global th =
blanchet@42641
   835
      if is_chained th then
blanchet@42641
   836
        Chained
blanchet@42641
   837
      else if global then
blanchet@42641
   838
        let val t = prop_of th in
blanchet@42641
   839
          if Termtab.defined intros t then Intro
blanchet@42641
   840
          else if Termtab.defined elims t then Elim
blanchet@42641
   841
          else if Termtab.defined simps (normalize_simp_prop t) then Simp
blanchet@42641
   842
          else General
blanchet@42641
   843
        end
blanchet@38937
   844
      else
blanchet@42641
   845
        if is_assum th then Assum else Local
blanchet@38738
   846
    fun is_good_unnamed_local th =
blanchet@38820
   847
      not (Thm.has_name_hint th) andalso
blanchet@38738
   848
      forall (fn (_, ths) => not (member Thm.eq_thm ths th)) named_locals
blanchet@38644
   849
    val unnamed_locals =
blanchet@38820
   850
      union Thm.eq_thm (Facts.props local_facts) chained_ths
blanchet@38820
   851
      |> filter is_good_unnamed_local |> map (pair "" o single)
blanchet@38627
   852
    val full_space =
blanchet@38738
   853
      Name_Space.merge (Facts.space_of global_facts, Facts.space_of local_facts)
blanchet@38752
   854
    fun add_facts global foldx facts =
blanchet@38699
   855
      foldx (fn (name0, ths) =>
blanchet@41989
   856
        if not really_all andalso name0 <> "" andalso
blanchet@39012
   857
           forall (not o member Thm.eq_thm add_ths) ths andalso
blanchet@38699
   858
           (Facts.is_concealed facts name0 orelse
blanchet@42728
   859
            (not (Config.get ctxt ignore_no_atp) andalso
blanchet@42728
   860
             is_package_def name0) orelse
blanchet@42728
   861
            exists (fn s => String.isSuffix s name0)
blanchet@42728
   862
                   (multi_base_blacklist ctxt) orelse
blanchet@38699
   863
            String.isSuffix "_def_raw" (* FIXME: crude hack *) name0) then
blanchet@38627
   864
          I
blanchet@38627
   865
        else
blanchet@38627
   866
          let
blanchet@38699
   867
            val multi = length ths > 1
blanchet@41279
   868
            val backquote_thm =
blanchet@41279
   869
              backquote o string_for_term ctxt o close_form o prop_of
blanchet@38699
   870
            fun check_thms a =
wenzelm@42361
   871
              case try (Proof_Context.get_thms ctxt) a of
blanchet@38699
   872
                NONE => false
blanchet@38699
   873
              | SOME ths' => Thm.eq_thms (ths, ths')
blanchet@38627
   874
          in
blanchet@38699
   875
            pair 1
blanchet@38699
   876
            #> fold (fn th => fn (j, rest) =>
blanchet@42641
   877
                        (j + 1,
blanchet@42641
   878
                         if is_theorem_bad_for_atps th andalso
blanchet@42641
   879
                            not (member Thm.eq_thm add_ths th) then
blanchet@42641
   880
                           rest
blanchet@42641
   881
                         else
blanchet@42641
   882
                           (((fn () =>
blanchet@42641
   883
                                 if name0 = "" then
blanchet@42641
   884
                                   th |> backquote_thm
blanchet@42641
   885
                                 else
blanchet@42641
   886
                                   [Facts.extern ctxt facts name0,
blanchet@42641
   887
                                    Name_Space.extern ctxt full_space name0,
blanchet@42641
   888
                                    name0]
blanchet@42641
   889
                                   |> find_first check_thms
blanchet@42641
   890
                                   |> (fn SOME name =>
blanchet@42641
   891
                                          make_name reserved multi j name
blanchet@42641
   892
                                        | NONE => "")),
blanchet@42641
   893
                              locality_of_theorem global th),
blanchet@42641
   894
                              (multi, th)) :: rest)) ths
blanchet@38699
   895
            #> snd
blanchet@38627
   896
          end)
blanchet@38644
   897
  in
blanchet@38752
   898
    [] |> add_facts false fold local_facts (unnamed_locals @ named_locals)
blanchet@38752
   899
       |> add_facts true Facts.fold_static global_facts global_facts
blanchet@38644
   900
  end
blanchet@38627
   901
blanchet@38627
   902
(* The single-name theorems go after the multiple-name ones, so that single
blanchet@38627
   903
   names are preferred when both are available. *)
blanchet@42728
   904
fun rearrange_facts ctxt only =
blanchet@38744
   905
  List.partition (fst o snd) #> op @ #> map (apsnd snd)
blanchet@42728
   906
  #> (not (Config.get ctxt ignore_no_atp) andalso not only)
blanchet@42728
   907
     ? filter_out (No_ATPs.member ctxt o snd)
blanchet@38627
   908
blanchet@41199
   909
fun external_frees t =
blanchet@41199
   910
  [] |> Term.add_frees t |> filter_out (can Name.dest_internal o fst)
blanchet@41199
   911
blanchet@42638
   912
fun relevant_facts ctxt (threshold0, threshold1) max_relevant is_built_in_const
blanchet@42638
   913
                   fudge (override as {add, only, ...}) chained_ths hyp_ts
blanchet@42638
   914
                   concl_t =
blanchet@37538
   915
  let
wenzelm@42361
   916
    val thy = Proof_Context.theory_of ctxt
blanchet@38822
   917
    val decay = Math.pow ((1.0 - threshold1) / (1.0 - threshold0),
blanchet@38822
   918
                          1.0 / Real.fromInt (max_relevant + 1))
blanchet@39012
   919
    val add_ths = Attrib.eval_thms ctxt add
blanchet@38696
   920
    val reserved = reserved_isar_keyword_table ()
blanchet@41199
   921
    val ind_stmt =
blanchet@41199
   922
      Logic.list_implies (hyp_ts |> filter_out (null o external_frees), concl_t)
blanchet@41211
   923
      |> Object_Logic.atomize_term thy
blanchet@41199
   924
    val ind_stmt_xs = external_frees ind_stmt
blanchet@40204
   925
    val facts =
blanchet@38699
   926
      (if only then
blanchet@38752
   927
         maps (map (fn ((name, loc), th) => ((K name, loc), (true, th)))
blanchet@40205
   928
               o fact_from_ref ctxt reserved chained_ths) add
blanchet@38699
   929
       else
blanchet@42641
   930
         all_facts ctxt reserved false add_ths chained_ths)
blanchet@42728
   931
      |> Config.get ctxt instantiate_inducts
blanchet@41273
   932
         ? maps (instantiate_if_induct_rule ctxt ind_stmt ind_stmt_xs)
blanchet@42728
   933
      |> rearrange_facts ctxt only
blanchet@38937
   934
      |> uniquify
blanchet@37538
   935
  in
blanchet@42646
   936
    trace_msg ctxt (fn () => "Considering " ^ string_of_int (length facts) ^
blanchet@42646
   937
                             " facts");
blanchet@39366
   938
    (if only orelse threshold1 < 0.0 then
blanchet@40204
   939
       facts
blanchet@39366
   940
     else if threshold0 > 1.0 orelse threshold0 > threshold1 orelse
blanchet@39366
   941
             max_relevant = 0 then
blanchet@38739
   942
       []
blanchet@38739
   943
     else
blanchet@41200
   944
       ((concl_t |> theory_constify fudge (Context.theory_name thy)) :: hyp_ts)
blanchet@41200
   945
       |> relevance_filter ctxt threshold0 decay max_relevant is_built_in_const
blanchet@41200
   946
                           fudge override facts)
blanchet@38822
   947
    |> map (apfst (apfst (fn f => f ())))
blanchet@37538
   948
  end
immler@30536
   949
paulson@15347
   950
end;